Play-Persona: Modeling Player Behaviour in Computer Games

7. RESULTS
This section presents the main findings of the clustering approaches applied to the data. A pre-
processing analysis of the data is complementary to and followed by the design of ESOM approach
for unsupervised learning of the data and the identification of the different player styles.
7.1 Pre-processing and Initial Cluster Analysis
All six features extracted are uniformly normalized into [0; 1] before any clustering analysis is
followed. Note that the cause of death features (Do, De and Df) are already normalized in [0; 1]
being percentages of the total number of deaths.
To get some first insight of the possible number of data clusters existent in the data, we apply the
k-means clustering algorithm to the normalized data for all k values less than or equal to 20. The
number of player observations (6- dimensional vector samples) and the sum of the Euclidean
distances between each player instance and its corresponding cluster centroid (quantization error)
are calculated for all 20 trials of the k-means algorithm. The analysis shows that the percent
decrease of the mean quantization error due to the increase of k is notably high when k = 3 and k
= 4. For k = 3 and k = 4 this value equals 19.06% and 13.11% respectively while it lies between
7% and 2% for k > 4. Thus, the k-means clustering analysis provides the first indication of the
existence of 3 or 4 main clusters within the data.
An alternative approach to k-means for cluster analysis is through hierarchical clustering. This
approach seeks to build a hierarchy of clusters existent in the data. The squared Euclidian distance
is used as a measure of dissimilarity between data vector pairs and Ward’s clustering method [20]
is utilized to specify the clusters in the data; the resulting dendrogram is depicted in Fig 3. (A
dendrogram is a treelike diagram that illustrates the merging of data sets into clusters. It consists
of many U-shaped lines connecting the clusters while the height of each U represents the squared
Euclidian distance between the two clusters being connected.)
Depending on where the designer sets the squared Euclidian distance threshold, T, a dissimilar
number of clusters can be observed. For instance, 3, 4 and 5 clusters of data can be identified if
6:56 > T > 4:72, 4:72 > T > 4:25 and 4:25 > T > 3:74, respectively.
Both clustering approaches demonstrate that the 1365 players’ feature vector can be clustered in a
low number of different player types. k-means statistics provide indications for 3 or 4 clusters while
the Ward’s dendrogram shows the existence of 2 populated and 2 smaller clusters, respectively, in
the middle and at the edges of the illustration resulting to four clusters. By further splitting the
216